Table 1.

Image based technique.

Ref	Year	Experimental Setup	Number of Subjects	Driving Duration	Type of Sensor	Physical Measurand	Detection Algorithm	Problem Solved	Limitations
[24]	2024		Public data set		Web Camera	Fatigue	CNN	The Cyber-Physical Systems are used to facilitate the real-time monitoring and analysis of the driving situation	It has a limited data set that will affect the model generality. The model must be validated with real-world challenges, such as lighting conditions.
[25]	2024	Driving simulator	Multiple drivers		Infrared cameras	Fatigue	Transfer learning with YOLOv8	The study proposed a fatigue-driving detection model that blends transfer learning methods with the YOLOv8n architecture.	The accuracy of fatigue identification can be impacted by individual variance because the model used in this work depends on factors like yawning and eye closure, which may not be relevant to all drivers.
[26]	2024		37 Subjects+ Public data set		Web Camera	Drowsiness	Deep Residual Networks (ResNet)	This study developed a deep learning architecture integrating residual and feature pyramid networks (FPN) to identify driver drowsiness.	Image quality, where the system's effectiveness relies on the quality of the input images. Additionally, the sensitivity of the camera position is essential.
[27]	2024		Public data set		Web Camera	Drowsiness	CNN and LSTM	The study combined IoT technology DL to develop an effective and unobtrusive drowsiness detection system. This integration allows real-time monitoring and notifications, which are critical for avoiding accidents.	Standard binarization algorithms struggle with dark skin tones, and limited variability in training data can affect their effectiveness. Testing conditions restricting rapid head movements may not accurately reflect real-world driving situations, impacting drowsiness detection.
[28]	2024		YAWDD dataset		Web Camera	Drowsiness	Deep Transfer Learning	The proposed method is intended to reduce false detections by employing ensemble learning and deep transfer learning models to ensure that only genuine drowsiness states are detected.	The accuracy of the DDD system depends on high-quality image processing. Factors like wearing sunglasses, changes in lighting, and camera-to-driver distance affect detection performance.
[29]	2024	Driving simulator	Public dataset		Web Camera	Drowsiness	lightweight convolutional neural networks	The suggested method employed lightweight CNN to ensure the system performed seamlessly on embedded devices such as Jetson Nano. This is essential for real-time applications.	The dataset has diverse photos; however, black-and-white images and varied lighting in the YALEB dataset can be challenging for lightweight models. The suggested network is lightweight but slightly more computationally complex than other networks.
[30]	2024	Driving simulator	37 subjects and two more online datasets		Web Camera	Drowsiness	two-branch multi-head attention architecture (TB-MHA)	The research presented a method for improving discrimination between various classes of samples (drowsy vs. awake) using a spatial filtering technique based on the Common Spatial Pattern (CSP) algorithm.	The technique detected fatigue mainly using facial features like landmarks and local areas. Still, its performance may be limited if facial visibility is compromised, such as when a driver wears sunglasses or has a partially concealed face.
[31]	2023	Actual driving	25 Subjects		RGB Camera	Drowsiness	Embedded system, edge computing, cloud computing modules	The study described an IoT-based automated approach for detecting Driver drowsiness This framework combined several components, such as an embedded platform, edge computing, and cloud computing, to give a solution for drowsiness detection and monitoring.	Reflections from drivers' spectacles can cause misclassifications and Reduced detection accuracy. Changing environmental conditions can also impact the system's performance, and there are potential data privacy risks with the cloud-hosted database for real-time monitoring.
[32]	2023	Driving simulator			Web Camera	Drowsiness	CNN model	The study proposed a method to detect driver fatigue by analyzing eye and mouth movements using a camera and a Convolutional Neural Network (CNN).	The current methods may fail at night or in poor conditions. Additionally, the driver's head posture can affect detection accuracy. The proposed model needs more real-world testing to assess the system's efficacy.
[33]	2023	Simulation	Multiple public data sets		Web Camera	Drowsiness	CNN model	This study examined the difficulty of achieving high precision on low-cost embedded devices like the Nvidia Jetson Nano. This makes the technology more practical for car use, improving safety.	The suggested method mainly detected signs of fatigue via the eyes and face. It did not look at other signs of fatigue, like posture and overall behavior, which could give a complete picture of a driver's condition.
[34]	2023	Actual driving	30 subjects		Web Camera	Yawning detection, fatigue detection	3D deep learning network and Bi-directional long short-term memory (LSTM)	This research addresses a few critical issues in yawning detection, particularly driver weariness, such as head posture variability, redundant frames, and lighting differences.	The detection accuracy depends on video data quality, including resolution, frame rate, and camera angle. It may also disregard minor yawning gestures between frames, resulting in false negatives.
[35]	2023	Driving simulator		1 h	Event camera	Yawning detection	lightweight deep learning models	It introduced event cameras, a neuromorphic sensing technology, for monitoring and verifying seatbelt fastening and unfastening. This approach used the unique capabilities of event cameras to give real-time analysis of dynamic behaviors that regular RGB or NIR cameras may not adequately capture.	The deployment of the proposed models within the limits of embedded hardware commonly seen in DMS is recognized as a difficulty. This constraint could impact the practical application of the system in actual automobiles, where computer resources may be limited.
[36]	2023	Simulator			Web Camera	Drowsiness And Fatigue	Threshold and DL based	The critical issue is the requirement for real-time monitoring of driver drowsiness. The suggested method provides instant feedback based on the driver's eye state, which is critical for prompt interventions to prevent accidents caused by drowsiness.	Drivers wearing sunglasses impair the model's effectiveness by preventing visual landmark identification and eye blink measurement, which are crucial for detecting fatigue. Facial landmark detection might also mislead the tracking driver if other objects block the face.
[37]	2022	Simulator	Multiple public data sets		Web Camera	Drowsiness	Various CNN models	The study introduced a real-time driver disturbance monitoring approach based on Convolutional Neural Networks (CNN). This technology is designed to assess driver drowsiness and fatigue, two crucial elements in road safety.	Implementing the proposed system in real-time applications may provide hurdles, notably in processing speed and computational requirements. The demand for high-performance hardware may limit technology accessibility in regular vehicles.
[38]	2022	Simulator	5 subjects		Camera	Fatigue	Machine learning and Resnet-50 models	This research aimed to develop a system for detecting driver fatigue by evaluating changes in facial features, particularly the eyes and mouth, in real-time.	The results showed that although some classifiers worked well, others did not. This variability indicates the system's performance may be inconsistent across different subjects or conditions.
[39]	2022	Actual driving	13 Subjects	1.5 h	Web Camera	Sleepiness	generic deep feature extraction module	The authors proposed personalized methods for identifying fatigue in drivers. They demonstrated the necessity of personalizing, as different people show signs of fatigue.	The authors acknowledged the limitations of their dataset, which was utilized to design and test the sleepiness detection system. They proposed that future research include independent datasets to validate these findings further.
[40]	2022		public data set		Web Camera	Drowsiness	two-stream spatial– temporal graph convolutional network (2s-STGCN)	This paper introduces the twin-stream spatial-temporal graph convolutional network (2s-STGCN). Many current driver drowsiness detection technologies face difficulties differentiating between various driving states, such as chatting, yawning, and blinking. This strategy sets out to rectify that.	The performance of the 2s-STGCN may deteriorate in challenging driving scenarios. Variations in lighting, shadows, and occlusions can all substantially impact the feature extraction process, potentially leading to misclassifications of driver states.
[41]	2022		public data set		Web Camera	Fatigue	multigranularity Deep Convolutional Model	The research proposed the Multi-Granularity Network (MEN), which used cues from partial face regions (such as the eyes, mouth, and glabella) to improve feature representation. This method addressed pose variability and enhanced the resilience of the feature extraction process.	The research acknowledged that the RF-DCM model cannot capture temporal information adequately. This limitation may impair the model's capacity to assess long-term relationships in time series data, critical for effectively diagnosing fatigue states over lengthy durations.
[42]	2022		public data set		Web Camera	Fatigue	(CNN)	This paper presented a systematic three-phase detection approach incorporating facial feature extraction, the Viola-Jones algorithm for identifying face traits, skin segmentation to ensure lighting invariance, and template matching for yawning and eye tracking.	This study revealed that drivers who wear glasses exhibit less precision in identifying facial characteristics and eye movements. The system's precision was inferior under low light conditions compared to daylight.
[43]	2021		synthetic event-based dataset		Neuromorphic vision sensors	Fatigue and drowsiness detection	(CNN)	The authors aimed to improve eye blink identification and analysis by developing a synthetic event-based dataset with accurate bounding box annotations, taking advantage of event cameras' high temporal resolution to improve overall driver safety and monitoring capabilities.	One crucial issue is the limited availability of event-based data, making it difficult to apply machine learning algorithms to event cameras successfully. The lack of data makes conducting thorough and statistically meaningful testing of the proposed approaches difficult.
[44]	2021	Simulator	8 Subjects		Web Camera	Fatigue	Detection fuzzy neural network	This study used an enhanced face identification approach to identify the driver's face in images collected by a CCD camera. It then used an ensemble of regression trees to identify face feature points, specifically the eyes and mouth.	The variation in recognition rates depending on head movement is one major limitation. Another significant issue with real-time face recognition systems is how varying lighting conditions or partially hidden facial features affect the system's effectiveness.
[45]	2021		Multiple Public data sets		Web Camera	Fatigue	CNN	In order to improve real-time performance and detection accuracy on edge computing devices, this research suggested a driver fatigue detection system that uses an optimized face alignment algorithm and a convolutional neural network.	The system's facial alignment performance may degrade under extreme conditions, such as when the driver is wearing sunglasses or in low-light situations. This can diminish detection accuracy, which is critical for real-world applications.
[46]	2021	Actual driving	10 Subjects	1 h	CCD Camera	Fatigue	Lightweight neural network model	In the context of vehicle-mounted embedded sensors with constrained memory and processing power, this paper attempted to address the problem of real-time driver fatigue detection using deep learning techniques on face video data.	Although the study claims to reach a detection speed of 27 FPS, more is needed in some real-time applications, particularly high-speed driving scenarios. Variations in the processing power of different vehicle-mounted embedded devices may alter the model's performance.
[47]	2021	Simulator	Public Data Set		HD camera	Yawning and Fatigue Detection	3D convolutional and BiLSTM networks	This paper presents a novel keyframe selection technique that reduces computational expenses and eliminates unnecessary frames from frame sequences. Achieving rapid detection of the most relevant frames is crucial for real-time processing.	The effect of low image quality on detection accuracy is one significant limitation. Additionally, significant camera vibrations might result in missing or false detections since face features are not always captured accurately.
[48]	2020	Simulator	Public data set		Web Camera	Drowsiness	A depth wise separable 3D convolutions	The study described a real-time method for detecting driver drowsiness using mobile platforms. The research emphasizes the usefulness of depth-wise separable 3D convolutions, which enable spatial and temporal data integration.	The method demanded significant processing capacity, notably for inference of 10-frame sequences. This high demand may limit the viability of real-time applications on lower-powered mobile devices. Because of these computational limits, the report suggests that other ways may be faster right now.
[49]	2020	Simulator	Public data set		Web Camera	Fatigue	Multi-task CNN model	The study suggested a Multi-task ConNN model using eye and mouth characteristics to measure driver fatigue. The model used the percentage of eye closure (PERCLOS) and the frequency of yawning as essential markers to determine the fatigue level of drivers.	The study used a constant frequency range and a set number of frames for analysis. This rigidity may limit the model's flexibility to diverse driving situations or surroundings, possibly affecting its real-world performance.
[50]	2020	Simulator	Public data set		Web Camera	Fatigue	a deep cascaded convolutional neural network (DCCNN	The research developed a Real-time and Robust Detection System (R2DS), to improve the accuracy, speed, and robustness of fatigue detection. This framework contains three main modules: facial feature extraction, ocular area extraction, and fatigue detection.	While the DCCNN is intended to increase detection accuracy, the study recognizes that methods based on artificial neural networks (ANN) frequently exhibit inadequate real-time performance due to their complicated structures and the requirement for considerable training data.
[51]	2019	Simulator	Public data sets		surveillance digital camera	Fatigue and Drowsiness	hybrid of CNN and (LSTM)	The research proposed the Eye Feature Vector (EFV) and Mouth Feature Vector (MFV) as evaluation parameters for determining the driver's eye and mouth states. These vectors are critical for estimating fatigue levels based on visual cues like eye closure and yawning.	The model detected fatigue primarily through facial indicators such as eye and mouth states. This concentration may neglect other signs of fatigue, such as physiological signals or behavioral patterns, which could provide a more comprehensive knowledge of a driver's status.
[52]	2019	Simulator	Public data sets		Web Camera	Drowsiness	A deep cascaded Convolutional neural network	The research described a condition-adaptive representation learning method for detecting driver drowsiness under various driving scenarios, including varied times of day and changes in the driver's appearance.	The proposed model requires a lot of labeled training data to cover various driver circumstances and scenarios appropriately. This can be a restriction when gathering such extensive data is unfeasible.
[53]	2019	Simulator	Public data sets		Web Camera	Drowsiness	convolutional control gate based recurrent neural network (ConvCGRNN)	The article developed a deep neural network (DNN) that can identify driver drowsiness in real time using video data. The network used CNN and ConvCGRNN, as well as a voting layer, to assess temporal relationships in facial data taken from videos.	The model's performance varies depending on the scenario, such as when drivers wear sunglasses or spectacles, which can obscure face features. Furthermore, the system depended substantially on consistent and precise facial tracking.
[54]	2019	Simulator	Public data sets		Web Camera	Drowsiness	Condition adaptive representation learning framework based on CNN	This approach combined four major models: spatiotemporal representation learning, scene condition understanding, feature fusion, and sleepiness detection. This system is intended to handle varied driving conditions in an adaptable manner and has been validated using the NTHU drowsy driver detection video dataset.	A significant limitation of this work is that the framework requires a large amount of labeled training data to handle different driving situations and scenarios adequately. Collecting such extensive data can be challenging.
[55]	2019	Simulator	10 subjects		Web Camera	Fatigue	Multiple Convolutional Neural Networks (CNN)-KCF (MC-KCF)	The system used a detection algorithm based on 68 critical facial locations, allowing for exact identification of facial regions crucial to drowsiness detection. This improved the system's accuracy in detecting indicators of weariness.	Drivers' heights can influence the location of their faces in the camera frame, leading to detection accuracy discrepancies. This diversity can make it challenging to ensure consistent performance for a broad group of users.
[56]	2018	Simulator	9 Subjects		Web Camera	Fatigue and drowsiness detection	Transfer learning classifier based on fast wavelet transforms and separator wavelet networks	The first essential contribution is the development of an eyes classifier that uses two transfer learning classifier designs. The second contribution is designing a fuzzy logic decision assistance system that divides driver vigilance into five categories.	Although the system is intended for real-time processing, the efficiency of the methods utilized (such as the Viola and Jones algorithm) can vary depending on the computational resources available. In resource-constrained contexts, this may impact system performance and dependability.